Research On Sensing Target Detection Algorithm Based On Feature Fusion And Attention Mechanism

Target detection refers to the detection of the categories and specific locations of objects of interest in an image given a given image.With the development of computer vision,object detection is increasingly widely used in daily life and work,and plays an important role in transportation,remote sensing,military,medical and other fields.In recent years,with the development of satellite remote sensing and aerial photography technology,the resolution and image quality of remote sensing images have become increasingly high,data sources have become increasingly rich,and detection difficulties have become increasingly difficult.Traditional target detection technology is based on artificially designed features,with low detection accuracy and low detection speed,which cannot meet the purpose of real-time high-precision detection of targets.At present,the application of deep learning in target detection has developed rapidly,and good results have been achieved in remote sensing target detection.Remote sensing images are affected by factors such as background and light,large differences in target scales,uneven target alignment directions,and large and densely distributed small size targets.Target detection accuracy is low,efficiency is poor,and it is prone to false detection,missed detection,and overlapping detection frames,which greatly affects detection work.Therefore,this article is based on target detection and improves the YOLOv4 detection algorithm.The main innovations are as follows:(1)A remote sensing target detection algorithm based on YOLOv4 improved feature fusion and global perceptual attention is proposed.An improved feature fusion module(P-bifpn)is used instead of PANet to increase cross scale connectivity while introducing weights at the output end,enhancing the expressiveness of important features,achieving efficient multiscale feature fusion,and resolving the accuracy degradation caused by multiscale changes.Then,a new global attention mechanism(GANet)is adopted to enhance the output of the sigmoid function while reducing average pooling and computational complexity,enhance the model’s learning of target context,capture interdependencies between targets,and reduce noise interference and global information loss.(2)Aiming at the small size and densely distributed features of remote sensing image targets,a visual transformer based MS transformer module is proposed,which is introduced into the feature fusion module of YOLOv4 to enhance the feature extraction ability of the detection algorithm.Because Transformer itself is a self attention mechanism,it obtains relevant information between targets through the self attention mechanism,enhancing the ability to detect dense targets.In view of the directional characteristics of the target arrangement in remote sensing images,this article improves the YOLOv4 target detection frame,using a five-coordinate representation to achieve multi-angle remote sensing target detection,and uses the smooth-L1-Io U loss function to make the prediction frame closer to the detected object.Aiming at the problem of overlapping prediction frames for dense targets,a soft NMS suppression method is adopted to further optimize the detection performance of the model.To verify the effectiveness of the algorithm in this article,experiments were conducted on remote sensing datasets RSOD and NWPU VHR-10 based on horizontal target detection,DOTA datasets based on rotating target detection,and VOC2007+2012 based on generalization experiments.The results were compared with relevant mainstream horizontal target detection algorithms and rotating target detection algorithms.The results show that the proposed algorithm can effectively deal with target detection of horizontal and rotating frames,effectively improving the detection ability of the model.